Dec 12, 2013
Trend 8: Explosion of Enterprise Flash
The enterprise flash storage market is maturing and enterprise customers are taking a closer look at SSD’s which are primarily designed for the client consumer market where manufacturing costs are the primary concern. The initial excitement over flash performance is becoming tempered by concerns over the “write cliff,” storage density, and durability that are the result of using commodity SSDs in enterprise storage systems.
SSDs have been one of the biggest trends in storage in 2013. Trend focus is predicting a 14% growth rate for SSDs in the enterprise market. However, the biggest growth is still predicted for the client market, 53%, where most SSD vendors are still focused. Since most of the revenue comes from the client market, SSD vendors are not motivated to redesign the SSD for enterprise storage use.
Flash NAND technology is not a random access technology for block oriented storage. It requires a programed erase before writes, which creates latency and address translation overhead. The data is written to pages that are contained with in a block. When data is rewritten, the old page is marked invalid and the data is written to a new page. Eventually the block fills up with invalid and valid pages and needs to be erased to reclaim the invalid pages. Before the block is erased the valid pages need to be stored in a buffer and saved until the block is erased before it can be rewritten. This process is known as garbage collection. During this garbage collection, write I/Os are blocked causing a degradation in writes, which is known as the “write cliff”
Flash is not really non-volatile. Electrons are leaking all the time and the voltage applied by reads, drives out more electrons. This requires a refresh when the signal degrades. Refreshes can be delayed with a longer ECC but that requires more processing power. A refresh means rewriting the page, which causes each write to be amplified and each write degrades the durability of the flash cells. The durability of MLC flash is 103 write/formats while the more expensive SLC flash is about 104 to 105. For comparison, hard disks have a durability of 1016 to 1018. Add to this workload, wear leveling, buffer capacity management and meta data processing and you see that an SSD requires a lot of processing power. Commodity SSDs, which are optimized for manufacturing costs, use a single core processor, which limits their ability to drive higher capacity SSDs. I/O performance suffers during garbage collection and durability suffers from write amplification.
Hitachi Data Systems has developed a controller chip for flash, which has the processing power to increase the durability, performance, and price capacity of MLC flash for enterprise storage. Using a quad core AMD processor, with 8 paths to the front and 32 paths to the flash DIMMS, this Flash Module Device, FMD, can manage 128 NAND chips and currently supports 3.2 TB with 40% higher performance than equivalent configurations of 400GB SSDs. The additional processor cores enable the garbage collection to be done outside of the I/O path to avoid the “write cliff” and the use of extended EEC to reduce write amplification. Real time write compression is also used to reduce the wear caused by writes and deliver up to 94% savings in storage space. For more information on the technology please see our white paper.
This FMD is available in in our VSP, HUS VM, and our HUS storage systems. HUS VM and HUS can be configured as an all-flash arrays with all the functionality of these storage controllers including in-system copies, replication, and tiering which is not generally available in all-flash arrays from other vendors. This support includes encryption in the back end directors (December 2013 for HUS 150), which enables the encryption of any device including FMDs and SSDs. The 1.6 TB and 3.2TB FMDs are available on VSP and HUS VM while the 1.6 TB FMD is available on HUS 150.
There are many startups developing storage systems using commodity SSDs and claiming to be enterprise storage systems. No matter what they do in the storage controller, the controller in the SSD will still have the limitations of a commodity SSD. Limitations like the “write cliff” may not be seen during initial acceptance testing, but will show up once all the spare blocks are used up.
Other enterprise storage vendors are beginning to follow our lead with larger capacity flash modules and the elimination of the “write cliff.” Flash technology has hit the peak of inflated expectations with SSDs. Now as enterprise users gain experience with flash technology we will see more focus on flash modules that are designed for enterprise storage and we will see an explosion of true enterprise flash in 2014.
See full list of my top ten trends for 2014 here.